The Research Of Bas-Reliefs Modelling Based On Learning Deformable 3D Models

The relief is carved on the plane,and it can express more information in a smaller space.According to the different height,relief can be divided into high relief,bas-relief and sunken relief.The bas-relief is most widely used in walls,utensils,and decorations.Traditional bas-relief is mainly carved by artists.With the rapid development of digital image and graphics,researchers have made great progress in the generation of three-dimensional(3D)digital relief.With the help of computer-aided,the generation of digital relief solves a series of problem of manual carving,which is skilled only,time-consuming,and not modifiable.Moreover,the generation of digital relief can make full use of existing images and three-dimensional models,it can quickly generate relief products which are easy to save.Because the convenience of using images as input in bas-reliefs generation is greater than using models as,in this paper,we focus on image-based bas-relief generation method,generate different bas-relief of objects of the same category and improve the efficiency.The main research content and conclusion of the paper are as follows:(1)Viewpoints estimation based on Non-Rigid Structure from Motion(NRSf M)Aiming at the ill-posed problem from image to 3D model,a camera viewpoint estimation method based on NRSf M algorithm is proposed.The method is feasible and low cost.It is not affected by the unknown camera parameters.And it is unstrict to dataset,so we can construct the dataset easily.Firstly,different key points are defined according to different structure,and then dataset are annotated and constructed.According to the two-dimensional image feature points sequence,the algorithm model is constructed,and the viewpoint and three-dimensional sparse shape of each input data are estimated finally,which is the foundation of learning deformation 3D models.(2)Deformable 3D model generation based on Visual HullBased on the result of viewpoints estimation,a method of learning deformation 3D models based on Visual Hull is proposed to solve the problem of 3D shape sparsity.Visual Hull is a surface reconstruction algorithm,which uses the contour information to reconstruct the model surface from the sparse point cloud shape and viewpoints.According to the key point,contour and direction of normal,we design constraints to adjust the 3D shape,and finally generate the deformable model.In this method,the similarity in objects of the same category is considered,and we can acquire prior knowledge from training data.The deformable model of this kind is reconstructed,and then provides the model basis for generation of new instances.The experimental results show that the method can reconstruct the deformable models of different categories and it is generalizability.(3)Bas-relief generation method based on deformable modelsBased on the results of viewpoints estimation and deformation 3D models,a bas-relief generation method based on deformation 3D models is proposed.The method based on deformation models is a top-down technology,which considers the psychological process and visual characteristics when people observe images.Inputting an image of a new instance,we can get a 3D model of the new instance according to the deformable model and deformation constraints.Compress the 3D model to get the shape of a bas-relief model of the instance.In order to correspond with the details extracted by the follow-up operations,loop subdivision is applied on the model to increase the number of vertices.However,the details in the image are missed.Taking an image as the input,firstly,the gray fusion gradient is used to enhance the edge information,and gamma correction is applied to enhance the details.Then,the effective light reflection model is used to reverse the depth value from the gray value.Finally,the detail information is superimposed on the instance model to generate the final bas-relief model.The results show that compared with model based,image gradient domain operation,and normal image-based methods,this method takes RGB image as the input,which is simple and efficient,and it is suitable for other types of bas-relief reconstruction.The reconstruction result reserves the shape and detail features of the original input image.The deformation 3D models in this method provides a new way for image-based bas-relief reconstruction.